Personalized multi-acoustic model training method, speech synthesis method and device

Abstract

A training method for multiple personalized acoustic models, and a voice synthesis method and device, for voice synthesis. The method comprises: training a reference acoustic model, based on first acoustic feature data of training voice data and first text annotation data corresponding to the training voice data (S11); acquiring voice data of a target user (S12); training a first target user acoustic model according to the reference acoustic model and the voice data (S13); generating second acoustic feature data of the first text annotation data, according to the first target user acoustic model and the first text annotation data (S14); and training a second target user acoustic model, based on the first text annotation data and the second acoustic feature data (S15).

Description

technical field [0001] The invention relates to the technical field of speech, in particular to a training method for a personalized multi-acoustic model for speech synthesis, a speech synthesis method and a device. Background technique [0002] Speech synthesis, also known as Text to Speech (Text to Speech) technology, is a technology that can convert text information into speech and read it aloud. It involves multiple disciplines such as acoustics, linguistics, digital signal processing, and computer science. It is a cutting-edge technology in the field of Chinese information processing. The main problem to be solved is how to convert text information into audible sound information. [0003] In the speech synthesis system, the process of converting text information into sound information is as follows: first, the input text needs to be processed, including preprocessing, word segmentation, part-of-speech tagging, polyphone prediction, prosodic level prediction, etc., and t...

AI Technical Summary

Problems solved by technology

Since two speakers generate the original speech according to different texts, and the pronunciation of the same syllable is obviously different in different sentence environments, therefore, if the same sound in different sentences of different speakers is made mapping, it is easy to cause the trained personalized acoustic model to be inaccurate, resulting in unnatural synthesized speech

[0010] For the second method, since the decision tree is a shallow model, its descriptive ability is limited, especially when the amount of user voice data is relatively small, the accuracy of the generated personalized acoustic model is not high, resulting in prediction The output parameters may be incoherent, which will cause jumps in the synthesized voice, unstable timbre, etc., resulting in unnatural voice

Images